Engram

Your OpenClaw agent is quietly forgetting everything you tell it. Decisions you made, preferences you stated, context from earlier in the session — gone. Not a bug. A structural limit of context windows.

Engram fixes this permanently.

🧠 WHAT IT DOES

Engram replaces OpenClaw's default sliding-window compaction with a DAG-based memory system. Every message is persisted in a local database. Nothing is thrown away. Your agent can reach back into weeks of conversation history and pull out exactly what it needs.

⚡ CORE FEATURES

→ DAG-Based Context Compaction — Messages stored in local SQLite, intelligently summarized into a directed acyclic graph. Summaries link back to sources. Raw messages always recoverable.

→ Pre-Compaction Fact Extraction — Before messages get summarized (a lossy step), Engram scans them for durable signals: decisions, preferences, entities, episodes. These facts survive forever, independent of summaries.

→ Entity Knowledge Graph — People, companies, projects, and decisions linked together with wiki-style connections. Your agent builds a growing map of everything it learns.

→ Cross-Agent Shared Memory — Multiple OpenClaw agents share one knowledge base. What one agent learns, every agent can recall.

→ Obsidian Vault Mirror — Your agent's entire memory becomes browsable Obsidian markdown notes. Search, link, and explore your agent's knowledge like your own second brain.

→ Intelligent Quality Gates — Not every sentence deserves to be remembered. Smart filtering ensures only high-signal facts get stored — no noise, no junk.

→ FTS5 Full-Text Search — Instant search across all memories, summaries, and raw messages. Find anything in milliseconds.

→ Built-In Agent Tools — lcm_grep, lcm_describe, lcm_expand, and memory_query let your agents search, inspect, and drill into their own history autonomously.

→ Cascade Condensation — Summaries of summaries keep context tight as conversations grow to thousands of messages. No manual intervention needed.

→ Conversation-Aware Organization — Memories are organized by conversation, session, and topic. Cross-conversation search works out of the box.

🔧 TECHNICAL DETAILS

→ SQLite-backed, local-first — no cloud dependencies, no API keys, no data leaving your machine

→ Works with any LLM provider configured in OpenClaw

→ 20+ tunable parameters: compaction thresholds, fresh tail size, summary model overrides, timezone, and more

→ Automatic migration system for seamless upgrades

→ 292 tests passing across 22 test files — production-ready from day one

→ Zero additional dependencies beyond OpenClaw itself

🎯 THE RESULT

It feels like talking to an agent that actually remembers you. Decisions you made last Tuesday, the file path you mentioned three sessions ago, that architectural choice from week one — all instantly recallable.

Stop repeating yourself to your agent. Install Engram.